Professional Documents
Culture Documents
A Study On Synergy of Multiple Crossover Operators in A Hierarchical Ga Applied To Structural Optimisation
A Study On Synergy of Multiple Crossover Operators in A Hierarchical Ga Applied To Structural Optimisation
DOI 10.1007/s00158-008-0268-x
RESEARCH PAPER
Received: 17 August 2006 / Revised: 15 December 2007 / Accepted: 2 April 2008 / Published online: 7 May 2008
© Springer-Verlag 2008
Abstract Development of crossover operators is based search and its individual properties have been investi-
on three different mechanisms: mating selection mech- gated extensively (Murata and Ishibuchi 1996; Spears
anism, offspring generation mechanism and offspring 1995; Hong et al. 1995; Aizawa et al. 1998; Eiben
selection mechanism. Most crossover operators are et al. 1998; Yoon and Moon 2002). Crossover operators
able to get exploration or exploitation of the domain generate new solutions by blending two current solu-
depending on the way they handle the current diversity tions. In general the action of the crossover operator
of the population. Each crossover operator directs the is complementary to the mutation operator producing
search towards a different region in the neighbourhood a synergy due to their different style of solution space
of the parents. The quality of the elements belonging traversal associated with crossover operators. Murata
to the visited region depends on the particular problem and Ishibuchi (1996) investigated the performance of
to be solved. This is confirmed by the well known No several crossover and mutation operators and they
Free Lunch (NFL) theorems. The simultaneous use observed that the combination of the best crossover
of diverse crossover operators on the population may and the best mutation did not give the best perfor-
induce more efficient algorithms. The aim of this paper mance among all combinations. The same researchers
is to analyse and to study complementary properties observed that there exist positive and negative effects
resulting from synergy effects using several crossover resulting from the combination of different crossovers
operators in particular for a hierarchical genetic al- and mutation operators.
gorithm. The reached improvements using multiple On the other hand, different crossover operators
crossover operators will be analysed through some have different behaviour over the solution space. Some
standard optimisation examples of hybrid composite researchers examined the synergy produced by combin-
structures. ing different styles of traversal of solution space. For
example Spears (1995) proposed an adaptive strategy
Keywords Hierarchical genetic algorithm · Multiple based on two different crossovers. Two-point crossover
crossovers · Synergy effects · Structural optimisation and uniform crossover were applied at a specific change
rate with benefit to the best-performing crossover op-
erator. Spears observed that the behaviour was inter-
1 Introduction mediate between the two GA using each individual
crossover operators without synergy effects. Never-
Crossover operator in Genetic Algorithms has been theless, Hong et al. (1995) using the same crossovers
referred as the most important operator in genetic but different strategies verified the synergy of multiple
crossovers. This fact shows that synergy depends on the
adopted strategy of combining the action of multiple
C. Conceição António (B)
Faculty of Engineering, University of Porto,
crossover operators during the evolutionary process.
4200-465, Porto, Portugal Several studies have revealed synergies using mul-
e-mail: cantonio@fe.up.pt tiple crossover operators (Aizawa et al. 1998; Eiben
118 C. Conceição António
et al. 1998; Yoon and Moon 2002). However, there adapted to work with real coding. In this later case
is not yet a relevant study in the literature covering the value of the gene corresponding to a position in
structural optimisation applications. This paper is a first the offspring is obtained by combining the values
study of synergy using multiple crossover operators of the genes of the parents in that same position
in the area of structural optimisation. A hierarchical (Herrera et al. 1998).
genetic algorithm is used to present the synergy effects – Offspring selection mechanism (OSM). Departing
of the multiple crossover operators’ action and how from the offspring generated for each set of parents
their different traversals on design space influence the this mechanism chooses the individuals that will be-
genetic search performance. come population members. One of the most widely
used OSMs chooses a core of best offspring to
form the next population (Wright 1991; Conceição
2 The crossover operator António and Lhate 2003).
Most of the crossover operators proposed in the
Early on combination was the most important task for literature is based on the generation of two offspring
developing the crossover operator, i.e., random parts chromosomes per each pair of parents. However, ap-
(schemata) of genetic material from two parents are proaches based on multi-parent crossover operators
combined to originate an offspring. From this point of (Kita and Kobayashi 1999; Tsutsui et al. 1999; Deb
view the goal of crossover operator is to share informa- et al. 2002) and crossover operators with multiple de-
tion between chromosomes. scendants (Wright 1991; Deb et al. 2002) have been
The above concept is a consequence of the orig- investigated. One offspring per each pair of parents is
inal Schema Theorem or the Fundamental Theorem considered in the present work. Any way, the OSM
of Genetic Algorithms proposed by Holland (1975) bounds the number of chromosomes from the offspring
where the focus is on implicit parallelism and ideal group to be inserted into the population.
sampling of schemata. Today the analysis based on
Scheme Theorem is overtaken due to the common use 2.2 Taxonomy analysis
of elitist strategies, heuristic rules, local optimisers and
different schemes of parent selection. This fact implies Following the taxonomy analysis presented by Herrera
the identification of recombination mechanisms and et al. (2003) the crossover operators based on applica-
a taxonomy analysis to understand the proposals for tion to two parents are grouped as:
designing crossover operators.
Discrete crossover operators (DCOs) A common
2.1 The mechanisms of recombination property of these crossover operators is that the value
of each gene in the offspring chromosome coincides
The exchange of structured data performed by with the value of the same gene in one of the parents.
the crossover operator is based on three different To obtain the gene values of the offspring from the
mechanisms: genes of parents there is no numerical transformation.
It considers the two-point and uniform crossover oper-
– Mating selection mechanism (MSM). This mech- ators (Syswerda 1989; Eshelman et al. 1989).
anism defines the process in which the chromo-
Aggregation based crossover operators (ABCOs) This
somes are mated before applying crossover on
category of crossovers groups operators that use an
them. Although, the most common procedure ran-
aggregation function to combine numerically the values
domly chooses the parents other approaches have
of the genes of the parents to generate the value of the
been proposed. For example, the elitist selection
offspring genes (Michallewicz et al. 1996).
(Conceição António and Lhate 2003) and the age-
based selection will be used in this work (Conceição Neighbourhood-based crossover operators (NBCOs)
António 2006). In this group of crossovers the offspring genes are
– Offspring generation mechanism (OGM). Produc- obtained from intervals defined in neighbourhoods
tion of new chromosomes from a set of parents associated with the genes of the parents throughout
selected by MSM is carried out by an appropriate probability distribution functions. Examples of NBCOs
recombination scheme. This mechanism enables are BLX-α, simulated binary crossover and fuzzy re-
the genetic material to be transferred from parents combination (Voigt et al. 1995), which are based on
to offspring. Different schemes have been proposed uniform, polynomial and triangular probability distri-
and all proposed OGMs for binary coding may be butions, respectively.
A study on synergy of multiple crossover operators 119
Hierarchical Genetic Algorithm with species Start migration: get the NA,2 best solutions
conservation from POP2 after isolation stage;
Initialisation of epochal evolution: Set e:=0 get the NA,3 best solutions from POP3 after
isolation stage;
Epoch(1) is the number of generations of isolation
Deletion of the worst solutions considering
stage for POP1
only the best fitted per species
for POP1 do
end do
Initialisation: random generation of the sub-
until Stopping criterion is met.
population considering one individual per species.
end do
Repeat
The evolution at isolation stage is performed ap-
for i=1 to Epoch(1) do (Isolation stage)
plying the genetic operators in the following order:
Evolve POP1; first selection, second crossover, third mutation, fourth
Only the best fitted of each species stays in elimination/replacement for species conservation para-
the sub-population; digm application.
end do The mutation operators used here are the implicit
Epoch(2) is the number of generations of mutation and controlled mutation and both of them
isolation are explained in Conceição António (2001, 2002, 2006).
stage for POP2 Implicit mutation considers a set of new chromo-
for POP2 do somes randomly generated and then inserted into the
population. Although those chromosomes exhibit a
Start migration: get the NA,2 best solutions from
probable fair fitness their effects on refreshing genetic
POP1;
material will emerge in the next generations through
Initialisation: random generation of the
combination with older chromosomes and so inducing
remaining sub-population.
population diversity. Conversely, controlled mutation
end do incorporates acquired data from the behaviour of the
for i=1 to Epoch(2) do (Isolation stage) state variables of the structural system into the selection
Evolve POP2; process of genes to mutate (Conceição António 2006).
Control the number of representative members Considering that a crossover operator influence
of a species, only the best fitted of each species study is the central investigation of this paper a detailed
stays in the sub-population. description will be presented in next section.
end do
Epoch(3) is the number of generations of
isolation stage for POP3 5 Proposal of crossover operators
for POP3 do
Start migration: get the best solutions from 5.1 Models overview
POP1 and POP2 rank them considering only
one per species and put the best NA,3 into The main objective of this section is to propose
POP3; Initialisation: random generation of the crossover operators of different style of design space
remaining sub-population. search to study the synergy effects. The crossover op-
end do erators used in the present study are:
for i=1 to Epoch(3) do (Isolation stage)
Evolve POP3; Elitist hybrid crossover with genetic improvement
Control the number of representative (EHCgi);
members of a species, only the best fitted of Elitist parameterised uniform crossover (EpUC);
each species stays in the sub-population. Age parameterised uniform crossover (ApUC).
end do Table 1 shows the mechanisms of recombination
for POP1 do of the proposed crossover operators grouping them
Set e := e + 1 according to taxonomy analysis. Due to the adopted
Get POP1 at epoch e to build the new sub- elitist strategy the first two crossovers, EHCgi and
population POP1; EpUC, exhibit exploitation properties. Since the mech-
anism of recombination is based on age without elitism
122 C. Conceição António
EHCgi Elitist based on fitness Hybrid crossover Elite group transferring NBCO
with genetic improvement
EpUC Elitist based on fitness Uniform parameterised crossover Elite group transferring DCO
ApUC Non-elitist based on age Uniform parameterised crossover Lethal age control DCO
+ updating access constraint
the last proposed crossover (ApUC) has exploration and the other one grouping the set with the
properties. worst fitness
The EHCgi operator is based on a mixed technique
L = S p+1t
, S p+2
t
, . . . , S Nt pop (2)
where besides considering the structured stochastic ex-
change of genetic material from the progenitors, the being Npop the dimension of the population.
local optimisation of genetic characteristic of the off- Step 3: Parent selection based on individual fitness.
spring is implemented. The couple of parents (p1 , p2 ) is obtained
From the taxonomy point of view this crossover using two independent selection processes in
operator is a neighbourhood-based crossover operator U and L sets. The parents p1 and p2 are
(NBCO). The two remaining crossovers are classified selected as
as discrete crossover operators (DCO).
p1 ∈ U and p2 ∈ L (3)
5.2 Mating selection mechanism
An individual Stk belonging to the U set has
The crossover operator transforms two chromosomes the following probability to be selected as
(parents) into a new chromosome (offspring), which parent:
genes come from both parents. Thus, the selection
t F IT Stk + F1t
process of parents plays an important role in the gen- P Sk = p
eration of offspring genetic material. In the proposed F IT Stk + F1t
crossover operators each pair of parents generates one k=1
The selection process is repeated until the necessary 5.3 Offspring Generation Mechanism (OGM)
couples (p1 , p2 ) (one per each offspring) are found.
Two OGMs are considered in the presented work: (1)
hybrid crossover with genetic improvement and (2)
5.2.2 Non-elitist based on age
uniform parameterised crossover.
A continuous model of generation of individuals was
adopted for age-structured population. An enlarged 5.3.1 Hybrid crossover with genetic improvement
population with age structure POP4 and performing
in parallel with the hierarchical topology of HGA is The hybrid crossover with genetic improvement is
considered in this model (Conceição António 2006). supported by commonality-based crossover framework
Each individual belonging to population POP4 is char- and it is based on heuristics linked to optimal design
acterized by two parameters: individual age and lethal of structures (Conceição António and Lhate 2003).
age. The individual age increases one unit after each According to the goal of genetic algorithms, informa-
generation. Any individual removed from HGA sub- tion from good parent solutions is exploited to find
populations either by elitist strategy or by finishing of better solutions. The concept of commonality-based
isolation stage of evolution and not selected for migra- crossover suggests that search must be driven in regions
tion, will survive in the population with age structure of neighbourhoods associated to parents and using a
POP4. Furthermore, its individual age will continue local optimiser.
increasing until removed definitively due to lethal age. The Hamming distance concept is used in the local
More details of the connection between age structure optimisation process to define a trajectory of search and
population POP4 and HGA populations are presented a local fitness is defined to rank the genetic material of
in Conceição António (2006). the offspring candidate (Conceição António and Lhate
In the population with age structure (POP4) the 2003). Considering the crossover of chromosomes or
MSM is conditioned by the candidate age. Assum- strings defined in the Hamming space Zl2 , any offspring
ing that population maturity and potentiality follow a generated by two parents is located, from the genetic
Normal distribution the parent selection is probability point of view, on a path that connects one parent to an-
dependent. Figure 2 presents the Normal probability other. This can be established introducing the concept
density function, fz (z), applied to parent selection in of intermediate vector defined as follows:
this crossover process. Individuals with ages located
at the tails of the normal density function are the Proposition The following definitions are equivalent,
youngest and the oldest of the scale, and they have a
very low probability to be selected as parents. Then the 1. The vector z defined in Zl2 is intermediate between
reproduction rate by crossover depends on the maturity two vectors v and y, written as v ♦ z ♦ y, if and
and degrades as the life cycle goes on till the end. only if
This dynamic behaviour is fundamental in the species
conservation context (Conceição António 2006). zi = vi or zi = yi , for i = 1 , . . . , n (8)
1
n-2
n-3
…
…
z
fz( z )
2 p
2
1 z age =1
Parent
age selection
124 C. Conceição António
2. Defining the distance of Hamming as the distance the offspring. This local optimisation performs on the
measured between two vectors in Zl2 given by neighbourhood defined by p1 and p2 considering the
associated Hamming space.
n
d (v, z) = |vi − zi |, (9) In the developed hybrid crossover scheme it is nec-
i=1 essary to define the local fitness function Φ(zk ). Since
the classical formulation of genetic algorithms only uses
z is intermediate between v and y if and only if
fitness information, it is necessary, in this case, to per-
d v, y = d (v, z) + d z, y (10) form the structural analysis for each solution. This can
be very tedious and computationally expensive, mainly
The above proposition clearly establishes that every in non-linear analysis. However, the use of approxi-
offspring generated by crossover is intermediate be- mations makes the evaluation process rational from
tween the parents not only in the genetic sense, but also the economical point of view. Thus the local fitness
in the sense of intermediate point of associated space, function is defined as
in this case the Hamming space.
The OGM of hybrid crossover with genetic improve- Φ(zk ) = β1 W(zk ) + β2 Ũ(zk ) + Ψ s (14)
ment combination scheme performs as follows:
where β 1 and β 2 are scaling constants, W(zk ) is the
1. Let p1 and p2 be two parents selected in an indepen- weight/cost of the structure, Ũ(zk ) is an approxima-
dent way, being p1 from the elite, U, defined in (1) tion
of the strain energy of the structural system and
and p2 from the group with less fitness, L, defined Ψ s is the constraint term related with the esti-
in (2). mated value of the constraint violation for the offspring
2. Define chromosome/solution zk intermediate to p1 candidate solution zk . The weight/cost of the struc-
and p2 , which genes of the active segments are ture is calculated directly using the decoding values
obtained as follows of the design variables. However, the strain energy and
the constraint term depend on the state variables of
p ⇐ p1 ∈ U i f Unif (0, 1) ≤ Picros
zk,i = 1,i (11) the structural system. Thus, the last two terms of (10)
p2,i ⇐ p2 ∈ L i f Unif (0, 1) > Picros
are evaluated based on approximations and heuristic
where zk,i denotes the string i of zk and Picros is the rules that can be found in Conceição António and
crossover probability. Lhate (2003).
3. Calculate the Hamming distances defined by
d1 = d p1 , zk and d2 = d p2 , zk , (12) 5.3.2 Parameterised uniform crossover
where only the active segments of the chromosome Using the known technique “parameterised uniform
are considered. crossover” proposed by Spears and DeJong (1991) the
4. The genes of a non-active segment of the chromo- offspring genetic material is obtained. This offspring
some/solution zk are equal to those of the corre- generation mechanism (OGM) performs a multipoint
sponding segment of the nearest parents, according combination of genes from both parents’ chromosomes
to the Hamming distances, d1 and d2 , calculated on (Conceição António 2001, 2002). The genetic material
the previous step. exchange can be implemented as follows:
5. Evaluate the local fitness function Φ(zk ) for the
chromosome /solution zk . 1. In active segments of the chromosome, the off-
6. Repeat from step 2 to 5 until a predefined num- spring gene zi is selected in a biased way given
ber N Z of chromosomes/solutions candidates are a probability Picros for choosing gene ui from the
obtained. progenitor chromosome that belongs to the elite
7. Calculate the local solution with best fitness, z, group U with best fitness defined in (1). This op-
defined as eration can be represented as
max (z) = Max (zk ) , k = 1, . . . , Nz (13) ui ⇐ S j ∈ U if Unif (0, 1) ≤ Picros
zi = (15)
8. Take the chromosome/solution z as the offspring vi ⇐ S j ∈ L if Unif (0, 1) > Picros
generated by p1 and p2 parents.
2. The genes of non-active segments of the offspring
The most important feature of the hybrid crossover chromosome are equal to the genes of the corre-
is the genetic material exchange of the parents based sponding segments in the chromosome of one of
on local search of the best genetic characteristic for the parents selected randomly.
A study on synergy of multiple crossover operators 125
5.4 Offspring Selection Mechanism (OSM) generation depending on the success rate of the genetic
operators, the access constraint is updated along the
The OSM is associated with the adopted strategy: (1) evolutionary process.
elite group transferring; (2) lethal age control plus up- The second aspect of this procedure is that an indi-
dating access. vidual assumes individual age equals to zero when it is
inserted in the age-structured population. The individ-
(a) Elite group transferring: ual age is increased by one unit at each generation and
the individual survives into the age-structured popula-
According to the evolutionary process for HGA sub- tion until it reaches the lethal age.
populations presented in Section 4 the offspring gener- In multiple crossovers the elitist strategy based on
ated by the crossover operator stays in the population fitness ranking elimination can coexist with the age-
for at least one generation. Since at each isolation stage based structure where the elimination of an individ-
the evolution is based on an elitist strategy where a core ual occurs only at lethal age. An individual with low
of best-fitted individuals is transferred from generation fitness can be selected as parent in subsequent gener-
into the next ones the OSM results in one of the follow- ations according to the species conservation paradigm
ing scenarios: (Conceição António 2006).
FIT zt ≤ FIT Stp : In this case the z offspring
is eliminated from the population at the (t + 1)-th 6 Strategies for multiple crossover applications
generation and only the best-fitted parent has some
probability of surviving
in the population; In this paper a study to examine the synergy effects
t resulting from the use of different crossovers is per-
FIT z > FIT S p : The offspring can integrate the
t
formed. Firstly it should be defined the concept of
elite group at the (t + 1)-th generation depending
synergy. The definition of synergy suggested by Yoon
on the fitness ranking. Offspring and/or best-fitted
and Moon (2002) is adopted as follows:
parent elimination can occur. The worst fitted parent
of the elite group is always eliminated.
Proposition Consider two crossovers C1 and C2 and
The survival time of offspring in the elite group of the assume without loss of generality that crossover C1 per-
population depends on the success rate of the genetic forms better than C2 when used alone. If the mixing of
operators to generate the best-fitted individuals. This C1 and C2 performs better than the sole usage of C1, we
parameter is used to measure the crossover perfor- say that crossovers C1 and C2 have synergy.
mance that will be defined further.
The adopted strategies to reach the synergy effects
(b) The updating access constraint and lethal age are based on the combination of proposed crossovers
control: applied to the isolation stage for each sub-populations
of the hierarchical genetic algorithm. Table 2 shows
The first aspect of this procedure is the integration six strategies using the proposed crossover operators
of any individual generated by “new” into the age- presented in Section 5. The first three strategies were
structured population denoted by POP4 (Conceição implemented taking advantage of exploitation charac-
António 2006). The access of individuals to this age- teristics of the elitist crossover. The last three strate-
structured population is regulated by a lower bound gies are based on mixed exploitation and exploration
constraint imposed on the fitness of the candidate. The properties of the elitism combined with age-structured
adopted strategy defines the lower bound as a percent- populations.
age of the
worst individual fitness of the elite group In the proposed strategies the synergy is reached
FIT Stp in HGA sub-populations at each generation using multiple crossover operators at the isolation stage
as follows: of HGA sub-populations and using heterogeneous dis-
tributed GA performed by hierarchical topology of
FITt+1 ≥ α FIT Stp (16) sub-populations.
access
From Table 2 it is observed the following functional
properties:
The access to age-structured population can be free
if α assumes an appropriate value
associated
for ex- – Crossover strategy 3 is a combination of crossover
ample with F2t in (7). Since FIT Stp changes at each strategies 1 and 2;
126 C. Conceição António
EHCgi elitist hybrid crossover with genetic improvement, EpUC elitist parameterised uniform crossover, ApUC age parameterised
uniform crossover with α = 0.6, ApUC* age parameterised uniform crossover with α = 0.95, []n sequence of multiple crossovers repeated
after n generations
– Crossover strategy 4 results from the combination longitudinal strength, X, transversal strength, Y, and
of crossover strategy 2 and the crossover ApUC shear strength, S. Other properties listed are longitu-
used at HGA sub-populations; dinal Young modulus, E1 , transversal elastic modulus,
– Crossover strategy 5 is a combination of crossover E2 , shear modulus, G12 , Poisson’s ratio, ν, and specific
strategies 2 and 4a; weight of the material, ρ. From Table 3 one mater-
ial is selected for each ply. The pair material/stacking
sequence defined by variable π j is a combination of
7 Results and discussion different materials for symmetric shell laminates. At
least two materials must be considered for hybrid com-
In order to show and to analyse synergy effects pro- posite laminate construction. The beam laminates have
duced using multiple crossovers two structural optimi- six plies made of material number 2 from Table 3 and
sation problems are presented. Selected parameters are this material does not change during the optimisation
used to evaluate the synergetic effects of the proposed process.
multiple crossover strategies: synergy, best fitness dif- The optimisation problem is formulated aiming the
ference (BFD) and crossover replacement rate in elite minimization of the weight/cost of the structure W(x,
group. These parameters will be introduced further in π) and the minimization of the strain energy U(x,
this section. π), subject to constraints related with structural in-
tegrity and with imposed service conditions as defined
7.1 Cylindrical shell problem in Conceição António (2002, 2006). The constraints are
imposed on the critical load factor and on the critical
7.1.1 Problem definition displacement, both of them associated with buckling
and first ply failure. The first ply failure is determined
A cylindrical shell made of laminated composite mate- based on Huber–Mises law (Conceição António 2002).
rials is considered as shown in Fig. 4. The shell is hinged Additionally displacement bounds can be imposed as
on straight sides and free on its curved boundary. A service conditions. Using an appropriate formulation
central point load Fmax 4 = 1 kN is applied and only for genetic search it is intended to maximise a global
a quarter of the structure was considered for the struc- fitness function FIT written as
tural optimisation problem. Ten laminates were taken
into account for the structure, four laminates grouping FIT = C1 − β1 W (x, π)
the shell elements (from 1 to 4) and six others (from 5
3
to 10) grouping the beam elements. All shell laminates −β2 U (x, π) − i [ϕi (x, π)] (17)
are symmetric and composed of six plies (Fig. 3). i=1
The mechanical properties of the materials for
ply laminate construction are presented on Table 3, with
⎧
⎨ 0 if ϕi (x, π) ≤ 0
i [ϕi (x, π)] = i = 1, 2, 3 (18)
⎩
Ki [ϕi (x, π )]qi if ϕi (x, π) > 0
A study on synergy of multiple crossover operators 127
m L
L1 R
θ θ
where i [ϕi (x, π)] represents each constraint term as- 2.0 × 10−2 m ≤ h j ≤ 4.0 × 10−2 m
sociated with constraint ϕ i (x, π), and the constants
5.0 × 10−3 m ≤ w j ≤ 1.5 × 10−2 m (19)
considered in (17) are C1 = 10, 000, β 1 = 10 and β 2 =
1000. Using previously acquired acknowledgement of Four digits are used for binary code format of ply
the structural behaviour, the constants β i are defined angle variables and three digits to codify each of the
aiming the numerical equilibrium between weight and remaining variables of the constraints (19). There are
energy terms of the fitness function (17). The constraint 24 possible combinations of three materials for the
term defined in (18) is calculated using penalties de- stacking sequence π j when considering six plies in the
pending on the constraint violation degree (Spears and symmetric jth composite shell laminate construction
DeJong 1991). The maximum allowed value for criti- and at least two materials per laminate.
cal displacement of buckling and for First Ply Failure All sub-populations (POP1, POP2 and POP3) of
displacement is da = 1.3 × 10−1 m. The same value is the hierarchical genetic algorithm have 15 individu-
taken for the allowable displacement at any point of the als. The elite group and the mutation group have five
structure. The lower bound for the critical load factor is individuals each in all sub-populations and three in-
λa = 0.45. The size constraints for the design variables dividuals participate in each migration flow between
are established as: the three HGA sub-populations. For each HGA sub-
population the isolation stage runs for six generations.
−90◦ ≤ θi, j ≤ 90◦ In age-structured population the lethal age is equal to
1.2 × 10−3 m ≤ ti, j ≤ 2.4 × 10−3 m 15 generations. The optimal solution for this problem
Table 3 Mechanical properties of the materials used for the composite laminate construction
Material number E1 [GPa] E2 (GPa) G12 (GPa) ν X (GPa) Y (GPa) S (GPa) ρ kg/m3
40
has been presented in a previous paper (Conceição crossover: 4a - 2 , synergy
crossover: 4a - 1 , BFD
António 2006).
20
synergy / BFD
7.1.2 Synergy analysis 0
-20
Synergy is measured by gains and losses in best fitness
using the proposition presented in Section 6 and can be
-40
written mathematically as
-60
Synergy = FITC1+C2 − FITC1 (20)
-80
where FIT C1 denotes the fitness obtained with the 0 12 24 36 48 60 72 84 96 108 120 132 144
when used alone and FITC1+C2 denotes the fitness ob- Fig. 5 Cylindrical shell example: synergy and BFD obtained
using crossover strategy 4a when compared to strategies 1
tained with the crossover strategy resulting from the and 2
mixing of C1 and C2.
In order to compare crossover strategies that do
not result directly from the application of the synergy Figure 4 shows the gains and the losses of the
definition it is introduced the concept of best fitness best fitness of the algorithm using crossover strategy 3
difference (BFD) as follows relatively to the best fitness obtained using crossover
strategy 1 and crossover strategy 2. Since the gains of
BFD = FIT S1 − FIT S2 (21) crossover strategy 3 are effective in both cases after the
20th generation the obtained results show the benefits
being S1 and S2 the two strategies under analysis. of using multiple crossovers. Only at the beginning
The idea is to identify the independent effects of of evolutionary process crossover strategy 3 does not
each crossover C1 and C2 and later to compare these have synergetic effects when compared to crossover
effects with those resulting from mixed use of both strategy 2.
crossover operators in multiple crossover strategy at Figure 5 shows the possible synergetic effect by us-
isolation stages of HGA sub-populations and in hetero- ing the age structure (POP4) according to crossover
geneous distributed GA with the hierarchical topology. strategy 4a (α = 0.6). Crossover strategy 4 results from
The first example shows the synergy that is obtained the mixed use of crossover strategy 2 and age para-
by mixing crossover strategies 1 and 2. Crossover strat- meterised uniform crossover (ApUC) at HGA sub-
egy 3 considers an alternate use of both crossover oper- populations. The ApUC is not elitist due the mating
ators elitist parameterised Uniform Crossover (EpUC) selection mechanism that is a function of individual
and elitist hybrid crossover with genetic improvement age as explained in Section 5. Since ApUC is based
(EHCgi) at sub-population POP2 level. on the species conservation paradigm its exploration
properties are good but its exploitation capacity is fair.
Then, crossover strategy 2 is better than the isolated
80
crossover: 3 - 2 usage of ApUC at HGA sub-populations in agreement
60 crossover: 3 - 1 with the comparison between crossover strategy 4a and
2 showed in Fig. 5.
40
According to the previous definition of synergy no
20 positive effects are identified during the first gener-
synergy
synergy
Considering that crossover strategy 5 results from Figures 8 and 9 show the improvements in best
the combination of crossover strategies 2 and 4a the fitness that result from the use of different strate-
synergy effects are observable in Figs. 6 and 7. The gies based on multiple crossovers and heterogeneous
synergy values displayed in Fig. 6 show that crossover distributed GAs. The BFD obtained from crossover
strategy 4b (α = 0.95) is better than crossover strategy strategies 3 and 4a proves that crossover strategy 3 is
4a (α = 0.6). Then it seems possible to increase the better than crossover strategy 4a as shown in Fig. 8.
synergy of crossover strategy 4 by increasing the value Figure 9 presents a similar comparison between
of the constant α on the updating access constraint. crossover strategy 3 and 5 using BFD values. Com-
From Fig. 7 the synergy effects of crossover strategy paring the values presented in Figs. 8 and 9 and since
5 are positive compared to crossover strategies 1 and 2. crossover strategy 5 is related to crossover strategy 4a
40
20
-20
-40
-60
0 12 24 36 48 60 72 84 96 108 120 132 144
generation
130 C. Conceição António
45
30
15
-15
-30
-45
0 11 22 33 44 55 67 78 89 100 111 122 133 144
generation
there is a clear improvement of the first one over the 7.1.3 Crossover replacement rate in elite group
last one. After 50 generations the difference between
crossover strategies 3 and 5 is not important. This This analysis parameter measures the success rate of
means that a synergy effect can be reached without replacement of at least one individual of the elite group
making use of pure elitism. by solutions obtained from crossover (offspring). The
Finally it can be concluded that synergy depends efficiency of the crossover strategy is analysed when the
on the frequency of interaction between different best offspring solution from crossover is compared with
crossover operators inside the HGA. The application of the worst fitted individual of the elite group. This last
numerical models based on multiple crossover usage in one is the best candidate to be eliminated in the next
GA is an important challenge for users and designers. generation. The replacement rate of the elite group is
60
45
30
15
–15
–30
–45
0 11 22 33 44 55 67 78 89 100 111 122 133 144
generation
A study on synergy of multiple crossover operators 131
70
a parameter associated with the quality of the solu-
crossover 5
tions obtained from crossover and it is defined after n
crossover 4b
generations as 60
n
event (i)
6
3
5
2
3
4 2 K
1
4 5 6 x
K
R1
three beam laminates (4 to 6) were considered, as For the synergy analysis the definitions of Section
defined on the left side of Fig. 12. 7.1.2 are adopted. It is intended to confirm the synergy
The problem full description presented in Sec- revealed in the previous example. Then, according to
tion 7.1.1 remains valid here. Exceptions are the pre- Table 2 the following strategies applied to spherical
scribed maximum displacement that will be da = 9.0 × shallow shell are analyzed:
10−2 m, and the constants in (17) that are β 1 = 100 and
– Crossover strategy 3 resulting from combination of
β 2 = 1. Since weight and energy assume different values
crossover strategies 1 and 2;
for this problem, the β 1 constants take new values. Size
– Crossover strategy 4b resulting from combination
constraints on design variables in (19) are imposed in
of crossover strategy 2 and the non-elitist crossover
the optimisation problem.
ApUC*.
The mechanical properties used in hybrid composite
laminate construction are presented in Table 3 and Figure 13 presents the synergy obtained using
the pair material/stacking sequence defined by variable crossover strategy 3. It can be concluded that the al-
π j is a combination of at least two different materials ternate use of elitist parameterised uniform crossover
at shell laminate level. The composite laminates are (EpUC) and elitist hybrid crossover with genetic im-
symmetric and have six plies each. The beam laminates provement (EHCgi) in crossover strategy 3 clearly pro-
are made of material number 2 from Table 3 and it duces a synergy over the results obtained from the
does not change along the optimisation procedure. The isolated use of EpUC in crossover strategy 1 and
objective is to maximize the fitness function defined in EHCgi in crossover strategy 2.
(17) and (18). The genetic parameters described at the Crossover strategy 4b (α = 0.95) and crossover
end of Section 7.1.1 are considered in the hierarchical strategy 2 presented in Table 2 are considered in
genetic algorithm solving this optimisation problem. the next comparison analysis. It should be reminded
A study on synergy of multiple crossover operators 133
synergy
25
15
5
-5
-15
-25
-35
generation
that crossover strategy 4b results from the mixed use result is a crossover strategy with higher performance
of crossover strategy 2 and age parameterised uniform as shown in Fig. 14. This confirms the previously made
crossover (ApUC) at HGA sub-populations. It must analysis for the cylindrical shell example. Furthermore
be noticed that strategy 2 is purely elitist and ApUC it is evident that it is possible to obtain synergy with-
is non-elitist. A fair exploitation capacity is expected out using purely elitist strategies. The combined and
for ApUC when compared with the elitists EpUC and simultaneous use of an elitist strategy and the species
EHCgi, both of them used in crossover strategy 2. conservation paradigm is possible though their appar-
The comparison of two different crossover strat- ent contradictory evolutionary concepts.
egies—4b and 2—shows the synergy obtained when The results obtained with this shallow shell example
a crossover with fair exploitation capacity as ApUC corroborate the previous study of synergy and rein-
is coupled with to an elitist crossover strategy. The force the idea that multiple crossover operators can be
30
20
synergy
10
-10
-20
-30
generation
134 C. Conceição António
explored to obtain the effectiveness of crossover between exploration and exploitation inducing im-
operator in GAs. portant synergetic effects.
– Hybridization is recommended as a promising
strategy to improve the effectiveness of crossover
8 Conclusions operators.
– It is possible to reach synergy by using crossover
This paper shows that efficient combination of multi- operators from different taxonomy groups. The
ple crossover operators in structural optimization can herein described combination of neighbourhood-
produce important synergy effects improving the per- based crossover operator (NBCO) together with
formance of Genetic Algorithms. In particular this discrete crossover operators (DCO) enhances this
concept is explored through the Hierarchical Genetic feature.
Algorithm (HGA) that results from application of the – The joint application of apparently contradictory
combination of hybrid crossover operators and hetero- evolutionary concepts such as pure elitism and
geneous distributed GAs. species conservation paradigm is possible. The
In order to study the synergy effects six cross- combination of elitist strategies together with age
over strategies are considered. These strategies were parameterised uniform crossover (ApUC) shows
built using several combinations of the following cross- synergetic improvements. However, some ability is
over operators: elitist hybrid crossover with genetic required so that complementarities between explo-
improvement (EHCgi), elitist parameterised uniform ration and exploitation of space design search can
crossover (EpUC) and age parameterised uniform be achieved.
crossover (ApUC). The use of multiple crossovers is
implemented and tested at both sub-population level Concluding, the use of multiple crossover operators
and hierarchical topology level. Considering two ex- and their hybridization is promising and necessary to
amples of structural optimisation different styles of obtain important synergies. Future research areas con-
exploitation and exploration of the design space of the cern the study of the synergy of multiple crossover
crossover strategies show different synergy effects as operators on problems with a design space of varied
follows: multimodalities.
third international conference on genetic algorithms. Mor- Proceedings of 4th annual conference on evolutionary pro-
gan Kaufmann, San Mateo, CA, pp 86–91 gramming. MIT Press, Cambridge, MA, USA, pp 367–384
Herrera F, Lozano M (2000) Gradual distributed real-coded ge- Spears WM, DeJong SK (1991) On the virtues of parameterized
netic algorithms. IEEE Trans Evol Comput 4:43–63 uniform crossover. In: Belew RK, Booker LB (eds) Pro-
Herrera F, Lozano M, Verdegay JL (1998) Tackling real-coded ceedings of the fourth international conference on genetic
genetic algorithms: operators and tools for behavioural algorithms. Morgan Kaufmann, San Mateo, CA, pp 230–236
analysis. Artif Intell Rev 12:265–319 Syswerda G (1989) Uniform crossover in genetic algorithms.
Herrera F, Lozano M, Sánchez AM (2003) A taxonomy for the In: Schaffer JD (ed) Proceedings of the third international
crossover operator for real-coded genetic algorithms. Int J conference on genetic algorithms. Morgan Kaufmann, San
Intell Syst 18:309–338 Mateo, CA, pp 2–9
Holland J (1975) Adaptation in natural and artificial system. The Tsutsui S, Yamamura M, Higuchi T (1999) Multi-parent re-
University of Michigan Press, Ann Arbor, MI combination with simplex crossover in real-coded genetic
Hong I, Kahng AB, Moon BR (1995) Exploiting synergies of algorithms. GECCO-99 Genetic and Evolutionary Com-
multiple crossovers: initial studies. IEEE conf. evolutionary putation Conference. Morgan Kaufmann, San Mateo, CA,
computation. Perth, Australia, pp 245–250 pp 657–664
Kita H, Kobayashi S (1999) Multi-parental extension of the uni- Tuson AL, Ross P (1998) Adapting operator settings in genetic
modal normal distribution crossover for real-coded genetic algorithms. Evol Comput 6:161–184
algorithms. International conference on evolutionary com- Voigt HM, Mühlenbein H, Cvetkovic D (1995) Fuzzy recombina-
putation’99. IEEE, Piscataway, NJ, pp 646–651 tion for the breeder genetic algorithm. In: Eshelman L (ed)
Michallewicz Z, Nazhiyath GG, Michallewicz M (1996) A note Proceedings of the sixth international conference on genetic
on usefulness of geometrical crossover for numerical opti- algorithms. Morgan Kaufmann, San Mateo, CA, pp 104–111
misation problems. Fifth annual conference on evolutionary Wolpert DH, Macready WG (1997) No free lunch theorems for
programming. MIT, Cambridge, pp 305–312 optimisation. IEEE Trans Evol Comput 1:67–82
Murata T, Ishibuchi H (1996) Positive and negative combination Wright A (1991) Genetic algorithms for real parameter optimiza-
effects of crossover and mutation operators in sequencing tion. In Rawlin GEF (ed) Foundations of genetic algorithms
problems. IEEE conf. evolutionary computation. Nagoya, 1. Morgan Kaufmann, San Mateo, CA, USA, pp 205–218
Japan, pp 170–175 Yoon H-S, Moon B-R (2002) An empirical study on the synergy
Spears WM (1995) Adapting crossover in evolutionary algo- of multiple crossover operators. IEEE Trans Evol Comput
rithms. In: McDonnell JR, Reynolds RG, Fogel DB (eds) 6:212–223